Re: Expensive memory.stat + cpu.stat reads

[Waiman Long <longman@redhat.com>]

On 7/13/23 19:25, Ivan Babrou wrote:
> On Mon, Jul 10, 2023 at 5:44 PM Waiman Long <longman@redhat.com> wrote:
>> On 7/10/23 19:21, Ivan Babrou wrote:
>>> On Wed, Jul 5, 2023 at 11:20 PM Shakeel Butt <shakeelb@google.com> wrote:
>>>> On Fri, Jun 30, 2023 at 04:22:28PM -0700, Ivan Babrou wrote:
>>>>> Hello,
>>>>>
>>>>> We're seeing CPU load issues with cgroup stats retrieval. I made a
>>>>> public gist with all the details, including the repro code (which
>>>>> unfortunately requires heavily loaded hardware) and some flamegraphs:
>>>>>
>>>>> * https://gist.github.com/bobrik/5ba58fb75a48620a1965026ad30a0a13
>>>>>
>>>>> I'll repeat the gist of that gist here. Our repro has the following
>>>>> output after a warm-up run:
>>>>>
>>>>> completed:  5.17s [manual / mem-stat + cpu-stat]
>>>>> completed:  5.59s [manual / cpu-stat + mem-stat]
>>>>> completed:  0.52s [manual / mem-stat]
>>>>> completed:  0.04s [manual / cpu-stat]
>>>>>
>>>>> The first two lines do effectively the following:
>>>>>
>>>>> for _ in $(seq 1 1000); do cat /sys/fs/cgroup/system.slice/memory.stat
>>>>> /sys/fs/cgroup/system.slice/cpu.stat > /dev/null
>>>>>
>>>>> The latter two are the same thing, but via two loops:
>>>>>
>>>>> for _ in $(seq 1 1000); do cat /sys/fs/cgroup/system.slice/cpu.stat >
>>>>> /dev/null; done
>>>>> for _ in $(seq 1 1000); do cat /sys/fs/cgroup/system.slice/memory.stat
>>>>>> /dev/null; done
>>>>> As you might've noticed from the output, splitting the loop into two
>>>>> makes the code run 10x faster. This isn't great, because most
>>>>> monitoring software likes to get all stats for one service before
>>>>> reading the stats for the next one, which maps to the slow and
>>>>> expensive way of doing this.
>>>>>
>>>>> We're running Linux v6.1 (the output is from v6.1.25) with no patches
>>>>> that touch the cgroup or mm subsystems, so you can assume vanilla
>>>>> kernel.
>>>>>
>>>>>   From the flamegraph it just looks like rstat flushing takes longer. I
>>>>> used the following flags on an AMD EPYC 7642 system (our usual pick
>>>>> cpu-clock was blaming spinlock irqrestore, which was questionable):
>>>>>
>>>>> perf -e cycles -g --call-graph fp -F 999 -- /tmp/repro
>>>>>
>>>>> Naturally, there are two questions that arise:
>>>>>
>>>>> * Is this expected (I guess not, but good to be sure)?
>>>>> * What can we do to make this better?
>>>>>
>>>>> I am happy to try out patches or to do some tracing to help understand
>>>>> this better.
>>>> Hi Ivan,
>>>>
>>>> Thanks a lot, as always, for reporting this. This is not expected and
>>>> should be fixed. Is the issue easy to repro or some specific workload or
>>>> high load/traffic is required? Can you repro this with the latest linus
>>>> tree? Also do you see any difference of root's cgroup.stat where this
>>>> issue happens vs good state?
>>> I'm afraid there's no easy way to reproduce. We see it from time to
>>> time in different locations. The one that I was looking at for the
>>> initial email does not reproduce it anymore:
>> My understanding of mem-stat and cpu-stat is that they are independent
>> of each other. In theory, reading one shouldn't affect the performance
>> of reading the others. Since you are doing mem-stat and cpu-stat reading
>> repetitively in a loop, it is likely that all the data are in the cache
>> most of the time resulting in very fast processing time. If it happens
>> that the specific memory location of mem-stat and cpu-stat data are such
>> that reading one will cause the other data to be flushed out of the
>> cache and have to be re-read from memory again, you could see
>> significant performance regression.
>>
>> It is one of the possible causes, but I may be wrong.
> Do you think it's somewhat similar to how iterating a matrix in rows
> is faster than in columns due to sequential vs random memory reads?
>
> * https://stackoverflow.com/q/9936132
> * https://en.wikipedia.org/wiki/Row-_and_column-major_order
> * https://en.wikipedia.org/wiki/Loop_interchange

Yes, it is similar to what is being described in those articles.


>
> I've had a similar suspicion and it would be good to confirm whether
> it's that or something else. I can probably collect perf counters for
> different runs, but I'm not sure which ones I'll need.
>
> In a similar vein, if we could come up with a tracepoint that would
> tell us the amount of work done (or any other relevant metric that
> would help) during rstat flushing, I can certainly collect that
> information as well for every reading combination.

The perf-c2c tool may be able to help. The data to look for is how often 
the data is from caches vs direct memory load/store.

Cheers,
Longman